Yuqi Meng1,2, Yihao Shi1,3, Kemeng Huang1,4, Zixuan Lu2, Ning Guo3, Taku Komura4, Yin Yang2, Minchen Li1,5
1Carnegie Mellon University 2University of Utah 3Zhejiang University 4The University of Hong Kong 5Genesis AI
ACM Transactions on Graphics (SIGGRAPH 2026)
We present an efficient B-spline finite element method (FEM) for cloth simulation. The pipeline combines B-spline-specific optimizations (reduced integration with split membrane/bending quadrature, accelerated Hessian assembly) with general numerical optimizations (a partial-factorization linear solver). This codebase is a standalone simulator that implements the B-spline FEM in the paper with IPC integration for contact handling.
If you find this work useful in your research, please cite our paper:
Yuqi Meng, Yihao Shi, Kemeng Huang, Zixuan Lu, Ning Guo, Taku Komura, Yin Yang, and Minchen Li. 2026. Efficient B-Spline Finite Elements for Cloth Simulation. ACM Trans. Graph. 45, 4, Article 102 (July 2026). https://doi.org/10.1145/3811278
@article{meng2026bspline,
author = {Meng, Yuqi and Shi, Yihao and Huang, Kemeng and Lu, Zixuan and Guo, Ning and Komura, Taku and Yang, Yin and Li, Minchen},
title = {Efficient B-Spline Finite Elements for Cloth Simulation},
year = {2026},
issue_date = {July 2026},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
volume = {45},
number = {4},
issn = {0730-0301},
url = {https://doi.org/10.1145/3811278},
doi = {10.1145/3811278},
journal = {ACM Trans. Graph.},
month = jul,
articleno = {102}
}BS-Cloth/
├── B-Spline-IPC/ # the simulator
│ ├── src/ # core C++ sources
│ │ ├── Core/, Geometry/ # fundamentals + B-spline / linear primitives
│ │ ├── Energy/ # membrane, bending, inertia, seam penalty
│ │ ├── Solver/ # Newton/IP solver, surface manager, config
│ │ └── Utility/ # CHOLMOD, OBJ exporter, YAML loader, math
│ ├── ext/ # vendored third-party libs
│ │ (Eigen, spdlog, fkyaml, stb, oneapi, IPC, ...)
│ ├── lib/, dll/ # prebuilt Windows libs / runtime DLLs
│ ├── shaders/ # GLSL shaders for the preview renderer
│ ├── models/ # cloth / garment / animation assets
│ ├── seam/ # seam definition files
│ ├── sample-configs/ # complex testcases presented in the paper
│ ├── config.yaml # active experiment config read at runtime
│ └── B-Spline-IPC.vcxproj # generated by premake
├── Altrar/ # in-tree Vulkan preview renderer
├── Prebuild/ # GenerateProject.bat / .sh
├── Execute/ # RunCurrentBuild.bat / .sh
├── webpage/ # project website
└── premake5.lua # workspace build script
Test assets live under B-Spline-IPC/models/ and B-Spline-IPC/seam/. Configs for the stress tests in the paper can be found in B-Spline-IPC/sample-configs/ references files. Details of configuration rules are elaborated in yaml-config-doc.md.
Most third-party libraries are vendored under B-Spline-IPC/ext/ (Eigen, spdlog, fkyaml, stb, oneAPI headers, IPC sources, …); on Windows, prebuilt binaries for SuiteSparse, LAPACK/OpenBLAS, and TBB are also shipped under B-Spline-IPC/lib/ with their runtime DLLs in B-Spline-IPC/dll/. The remaining dependencies must be installed by the user.
-
C++20 toolchain — Visual Studio 2022 (Windows), or GCC ≥ 11 / Clang ≥ 14 (Linux).
-
premake5 (≥ 5.0.0-beta8) — generates the project / makefiles. Available pre-built from the official website.
-
Intel oneAPI Base Toolkit — provides MKL (used as the Pardiso direct solver and as Eigen's BLAS/LAPACK backend) together with the Intel OpenMP runtime. The build script reads two environment variables to locate them:
MKLROOT— root of the MKL install (e.g.C:/Program Files (x86)/Intel/oneAPI/mkl/lateston Windows).CMPLR_ROOT— root of the Intel compiler install (e.g.C:/Program Files (x86)/Intel/oneAPI/compiler/latest).
Pass
--no-mkltopremake5to fall back to Eigen's native solver and skip MKL altogether (the env vars are then not required, at the cost of solver performance). -
Vulkan SDK — required by the in-tree preview renderer, which is enabled by default on Windows. Not needed if you build with
BSIPC_DISABLE_RENDERERdefined. -
Intel oneAPI TBB — set
TBBROOTto its install root so premake can find the TBB libraries. -
SuiteSparse (CHOLMOD, AMD, CAMD, COLAMD, CCOLAMD, suitesparseconfig) — install via your package manager, e.g.
sudo apt install libsuitesparse-dev. If high performance is desired, consider building from scratch and linking Intel MKL.
This project adopts premake as its build system, since its portability best suites the scale of the codebase.
To run the basic functionalities of this codebase on Windows, you need to have Visual Studio (preferably 2022) installed. Run Prebuild/GenerateProject.bat will generate the .sln Visual Studio solution files for you.
Premake does not have its latest version available on various package managers, so installing from its official website is required. Run the following command to install:
wget https://github.com/premake/premake-core/releases/download/v5.0.0-beta8/premake-5.0.0-beta8-linux.tar.gz
# the following should yield a premake5 executable in the same directory
tar -xvf premake-5.0.0-beta8-linux.tar.gz
# This should give something like "premake5 (Premake Build Script Generator) 5.0.0-beta8"
premake5 --versionConsider moving it to your binary directory (something similar to /usr/bin) for future usage.
To generate the makefiles, run Prebuild/GenerateProject.sh. Alternatively, go to the root directory of the project and run
premake5 gmake2
Add the flag --cc=clang if you are using clang family compilers. Check more customization on the official documentation.
There are various macros in the project, which gives fine control over toggling on/off of different functionalities.
Definition of BSIPC_DISABLE_RENDERER suppresses real-time rendering of the resulting garment along with the simulation. The renderer is rather simple and only serves as a preview of the result.
If this macro is disabled, the program requires installation of Vulkan Toolkit for rendering, as the simple renderer relies on Vulkan. BSIPC_DISABLE_RENDERER is on default disabled on Windows, and enabled on Linux.
The simulator records all the intermediate results of simulation using .yaml files in steps/ folder at the same location as the executable. Enabling BSIPC_YAML2OBJ will disable all the simulation, and instead creates a series of .obj file in the objs/ folder. This can then be used for visualization and rendering with professional renderer, for example Cycles in Blender. Importing can be done using Blender Sequence Loader.
The simulator reads configuration of experiments from config.yaml in the same position as the executable. It creates a log.txt logging concise information of the current status, and a steps/ folder creating a series of .yaml files saving the configuration of primitive at each step of simulation. If screenshot field is set to true in config.yaml and BSIPC_DISABLE_RENDERER is disabled, there will be an additional cache/ folder saving the rendering result at each step.
For simplicity of parameter setting, when quadrature scheme falls under the class of local quadrature, the per-patch quadrature order is hardcoded to be 2 for simplicity.
The codebase further enables resuming simulation from a given step, as long as the corresponding configuration is provided. To resume a simulation, copy the yaml file at that step, rename it to config.yaml, and ensure that continue_on_step field is set to be true (this will automatically be the case if you are using the yaml file generated by the codebase). Running the executable now will start simulation continuing on the designated step.